Omni-modal large language models (Omni-LLMs) extend traditional LLMs to handle multiple input types at once, including vision, audio, and inertial measurement unit (IMU) signals, alongside text. According to the paper's abstract, these models are designed to support a broader range of multimodal tasks than earlier systems.

The paper, posted on arXiv, opens with a claim that the authors 'find that these models…' before the abstract cuts off. The title, 'Omni-Modal LLMs Know More Than They Say,' suggests the central result is a mismatch between what the models have learned and what they output. However, the specific evidence and methodology are not visible in the available excerpt.

Because the source is a single truncated abstract, this digest can only report the paper's stated scope and its apparent thesis. Readers should consult the full paper for the actual findings.